Global Self-Supervised Graph Learning for Recommendation

In recent years, Comparative Learning (CL) has driven a series of effective research in the recommendation domain, and it excels in dealing with the data sparsity problem in recommender systems since no data labeling is required in CL. The typical model flow of the approach consists of structurally perturbed augmentation of the user-item bipartite graph, and model optimization by maximizing the consistency of node representations across different graph augmentations. Although the model has been shown to be effective, we believe that existing contrastive learning research and applications are not yet sufficient. To enhance the global nature of contrast learning, we introduce global graph view self-supervised graph learning (GSGL). This method adds an auxiliary self-supervised task to the traditional recommendation supervision task to realize multi-view contrast learning. Specifically, each node builds enhanced views by three data enhancement methods (node loss, edge loss, and random noise) and performs two-by-two comparison learning on these views. Through empirical studies on three benchmark datasets, the effectiveness of the GSGL algorithm is demonstrated, which not only leads to a more uniform view embedding spatial distribution and improves the recommendation accuracy, but also exhibits strong robustness to interaction noise. The code and used datasets are released at https://github.com/LX-Moon/GSGL.